$\textbf{Sonification 3.1}$ Sonification of 50 gradients. In this sonification method, all elements of a gradient were converted to a sine-wave. All the gradient-element sine-waves were then overlaid to sonify a gradient.

$\textbf{Sonification 3.2.1}$ C-4 played on the piano with velocity 20

$\textbf{Sonification 3.2.2}$ C-4 played on the piano with velocity 100

$\textbf{Sonification 3.4}$ 60 Histograms of 100 samples each drawn from B(9,9) and corresponding sonification.

$\textbf{Sonification 3.5}$ Sonifiation of two random variabes. Again, 100 samples were drawn from B(9,9) 60 times for both random variables.

$\textbf{Sonification 3.5.1}$ Only the first variable sonified by the piano

$\textbf{Sonification 3.5.2}$ Only the second variable sonified by the cello

$\textbf{Sonification 3.6}$ Sonifiation of two random variabes. Again, 100 samples were drawn from B(9,9) 60 times for both random variables. this time the left variable was sonified through a bass with low pitch. The right right was sonified through a pinao with high pitch. Addtionally, the Bass was only played on the left channel, and the piano on the right channel.

$\textbf{Sonification 3.6.1}$ Sonification of only the first variable with the bass

$\textbf{Sonification 3.6.2}$ Sonification of only the second variable with the piano

$\textbf{Sonification 3.6.3}$ Sonification of both variables without separation of sound sources

$\textbf{Sonification 4.1}$ Sonification of gradient elements, all elements included, a tuned learning rate of 2e-2 was used

$\textbf{Sonification 4.2}$ Sonification of gradient elements, with a better scaled x-axis.

$\textbf{Sonification 4.3}$ Sonification of gradient elements, with $log(\vert{}x_i\vert{}$) being applied to every element $x_i$ of the gradient. A tuned learning rate of 2e-2 was chosen

$\textbf{Sonification 4.3.h}$ Sonification of gradient elements, with $log(\vert{}x_i\vert{}$) being applied to every element $x_i$ of the gradient. Heatmap version of Sonification 4.3

$\textbf{Sonification 4.4}$ Sonification of gradient elements, with $log(\vert{}x_i\vert{}$) being applied to every element $x_i$ of the gradient. A small learning rate of 5e-5 was chosen as hyperparamter

$\textbf{Sonification 4.4.h}$ Sonification of gradient elements, with $log(\vert{}x_i\vert{}$) being applied to every element $x_i$ of the gradient. A small learning rate of 5e-5 was chosen as hyperparamter. Heatmap version of Sonification 4.4.

$\textbf{Sonification 4.5}$ Sonification of a gradient with a learning rate of lr4.8e-1

$\textbf{Sonification 4.5.h}$ Sonification of a gradient with a learning rate of lr4.8e-1, heatmap version

$\textbf{Sonification 4.6}$ Sonification of a training process with some vanishing gradients, loss also sonified

$\textbf{Sonification 4.7}$ Sonification of a gradients, with exploding gradients at the end of the sonification. loss also sonified

$\textbf{Sonification 4.8}$ Sonification of scalar products of the gradients within a mini-batch. The full range [-1,1] is sonified with one instrument, the histogram has 20 bins

$\textbf{Sonification 4.9}$ Sonification of scalar products of the gradients within a mini-batch. The range [-1,0] is sonified with a bass the range [0,1] is sonified with a piano. Furthermore, the bass is low pitched and only played on the left speaker, the piano is high pitched and only played on the right speaker. The histogram has 24 bins

$\textbf{Sonification 4.10}$ Sonification of scalar products of the gradients within a mini-batch. Only the range [0,1] is sonfied with a piano. The histogram has 20 bins

$\textbf{Sonification 4.10.h}$ Sonification of scalar products of the gradients within a mini-batch. Only the range [0,1] is sonfied with a piano. The histogram has 20 bins. Heatmap version

$\textbf{Sonification 4.11}$ Sonification of scalar products of the gradients within a mini-batch. Only the range [-1,0] is sonfied with a piano. The histogram has 20 bins

$\textbf{Sonification 4.12}$ trivial sonification of eigenvalue distribution, lr2e-2

$\textbf{Sonification 4.13}$ Sonification of eigenvalues for learning rate 2e-2. eigenvalues were separated into 2 Histograms. The left histograms covers the range [0,1] and is only played on the left speaker with a bass. The right histogram covers the range [1,12] and is played on the right speaker with a piano.

$\textbf{Sonification 4.14}$ Sonification of the top 10 eigenvalues per mini-batch

$\textbf{Sonification 4.14.h}$ Sonification of the top 10 eigenvalues per mini-batch. Heatmap version

$\textbf{Sonification 4.15}$ Sonification of $(\lambda_k,SNR(\lambda_k))$. $\lambda_k$ is sonified with a low pitched bass and played on the left speaker, $SNR(\lambda_k)$ is sonified with a high pitched piano and played on the right speaker.

$\textbf{Sonification 4.16}$ Sonification of $(\lambda_k,SNR(\gamma_k))$. $\lambda_k$ is sonified with a low pitched bass and played on the left speaker, $SNR(\gamma_k)$ is sonified with a high pitched piano and played on the right speaker.